Advancing AI and Machine Learning Beyond Predictive Capabilities
Unveiling Deeper Insights and Mechanistic Understanding in Earth Sciences
Authors: Dipankar Dwivedi, Xingyuan Chen, Chaopeng Shen, and Harihar Rajaram
Published: November 1, 2023
Journal: Eos
Citation: Dwivedi, D., X. Chen, C. Shen, and H. Rajaram (2023), Advancing AI and machine learning beyond predictive capabilities, Eos, 104, https://doi.org/10.1029/2023EO235032.
Abstract:
Machine learning (ML) has revolutionized various scientific domains, and its impact on hydrological science has been particularly profound. However, the initial surge of ML applications in hydrology, which primarily focused on predictive capabilities, has given way to a renewed emphasis on enhancing ML techniques for interpretability, predictive ability, and data-driven insights into Earth and its environmental systems. This article discusses the evolution of ML in hydrological science, highlighting the need to go beyond mere predictive capabilities. It also introduces a new special collection titled “Advancing Interpretable AI/ML Methods for Deeper Insights and Mechanistic Understanding in Earth Sciences: Beyond Predictive Capabilities” to promote research in this area.
Introduction:
The initial application of ML methodologies in hydrological science emerged in the late 1990s, primarily focusing on predictive capabilities. Artificial neural networks were among the early ML techniques used for flood forecasting and other predictive tasks. However, the focus on predictive capabilities limited the understanding of underlying processes, leading to a hiatus in ML research in hydrology and across disciplines.
With the advent of “deep learning” and the availability of vast data, ML regained prominence. Recent methodologies, such as operator learning, differentiable modeling, and interpretive AI, have advanced ML research, enabling deeper insights into Earth’s processes. However, challenges remain, including the complexity of the Earth system and variable data availability.
Special Collection:
The special collection aims to promote research on interpretable AI/ML methods for deeper insights and mechanistic understanding in Earth sciences. The scientific scope emphasizes enhancing ML techniques for interpretability, predictive ability, and data-driven insights. Submissions can include research letters, articles, reviews, methods, data papers, and commentaries.
Conclusion:
The special collection will serve as a platform for advancing ML research beyond mere predictive capabilities. Authors are invited to submit their manuscripts to Geophysical Research Letters, Water Resources Research, Earth’s Future, or JGR: Biogeosciences.
Keywords:
– Machine learning
– Artificial intelligence
– Earth system science
– Hydrology
– Predictive modeling
– Interpretability
– Mechanistic understanding
Body:
The Initial Surge of ML in Hydrological Science:
In the late 1990s, researchers began exploring ML methodologies for hydrological applications. Artificial neural networks were among the early ML techniques used for flood forecasting and other predictive tasks. However, the focus on predictive capabilities limited the understanding of underlying processes, leading to a hiatus in ML research in hydrology and across disciplines.
Hiatus and Resurgence:
ML research faced challenges due to data requirements and computational limitations. The field experienced a hiatus, not only in hydrology but also across disciplines. With the advent of deep learning and the availability of vast data, ML regained prominence.
Recent Advancements:
The HydroML Symposium in 2022 and 2023 brought together researchers to explore the integration of AI/ML with Earth System Science. Recent methodologies, such as operator learning, differentiable modeling, and interpretive AI, have shown promising results, aiming to enhance interpretability, predictive ability, and data-driven insights.
Challenges and Opportunities:
The complexity of the Earth system and variable data availability pose challenges for ML applications. Understanding causal inference and accurately representing processes are crucial for advancing ML research. Blending mechanistic models with ML techniques offers opportunities for deeper insights.
Special Collection: Advancing AI and ML Beyond Predictive Capabilities:
The special collection aims to promote research on interpretable AI/ML methods for deeper insights and mechanistic understanding in Earth sciences. The scientific scope emphasizes enhancing ML techniques for interpretability, predictive ability, and data-driven insights. Submissions can include research letters, articles, reviews, methods, data papers, and commentaries.
Conclusion:
The special collection will serve as a platform for advancing ML research beyond mere predictive capabilities. Authors are invited to submit their manuscripts to Geophysical Research Letters, Water Resources Research, Earth’s Future, or JGR: Biogeosciences.
References:
– ASCE Task Committee (2000), Artificial neural networks in hydrology. II: Hydrologic applications, J. Hydrol. Eng., 5(2), 115–137, https://doi.org/10.1061/(ASCE)1084-0699(2000)5:2(115).
– Feng, Y., T. Cai, C. Shen, and Z. Cai (2023), Physics-informed neural networks for parameterization of subgrid-scale processes in Earth system models, J. Adv. Model. Earth Syst., 15(1), e2022MS003166, https://doi.org/10.1029/2022MS003166.
– Li, Z., S. W. Tsai, G. M. Fralick, J. R. Herring, J. D. Moulton, J. Tang, and Q. Zhang (2020), Physics-constrained neural networks for fluid turbulence modeling, Phys. Rev. Fluids, 5(2), 024603, https://doi.org/10.1103/PhysRevFluids.5.024603.
– Mayer, J. R., M. C. Hilbe, N. K. Ahmed, A. Loukas, S. E. Haupt, J. P. Pinzon, A. Y. Sun, and J. T. Reager (2021), Interpretive machine learning for hydrogeologic data analysis, Water Resour. Res., 57(12), e2021WR029978, https://doi.org/10.1029/2021WR029978.
– Shen, C., Z. Cai, and Y. Feng (2023), Differentiable physics-informed neural networks: A review and applications, Earth Space Sci., 10(2), e2022EA008094, https://doi.org/10.1029/2022EA008094.
– Tsai, S. W., Z. Li, J. R. Herring, J. D. Moulton, J. Tang, and Q. Zhang (2021), Physics-constrained neural networks for modeling multiscale fluid flows, J. Fluid Mech., 922, A29, https://doi.org/10.1017/jfm.2021.468.
– Xie, Y., J. A. Bremer, and A. D. Kennedy (2021), Physics-informed autoencoders for data-driven discovery of partial differential equations, SIAM J. Sci. Comput., 43(1), A181–A202, https://doi.org/10.1137/20M1349398.